[computer-go] State of the art of pattern matching

[Mark Boon]

I did an experiment that looks rather similar. I generated patterns  
and only kept the ones that had a minimum amount of 'urgency' and a  
minimum number of occurrences. But I noticed two things when using  
these patterns in a MC playout:

1) There are many important moves missing. Apparently they were not  
picked up from the game-database even though the number of games is  
in the 10-thousands.

2) When I used these patterns during simulation the playouts suddenly  
look surprisingly like normal Go compared to random playouts.  
However, the program started to play worse. Much worse. Even when I  
let it compute as many playouts as a program without patterns.

The first observation made me wary to rely on harvested patterns. I  
think it shows at least it needs to be used in combination with hand- 
crafted patterns. Also it means that the fact you harvest a large  
number of patterns isn't necessarily meaningful.

The second observation I think may have been caused by not enough  
randomness. But that means I first have to find an answer to how much  
randomness I need to put into the patterns. I'm first looking at this  
question with some hand-crafted patterns to get a better handle on  
this issue.

Mark


On 30-mrt-08, at 09:14, Jacques Basaldúa wrote:

> Mark Boon wrote:
>
>
> >There are 16 4-distance points, so if you spill ino that by one
> > point you get 315 or a little over 14 million patterns. Multiplied
> > by 3 for every don't-care point within less than 4 distance. Ouch.
>
> True, but the number of patterns is learned automatically. When I
> first learn the 55K+ games, there are so many patterns that I can
> only create a pattern file of less than 2000 games. I create 32
> such files an call "importance" the number of files in which each
> pattern is found. (a value from 1 to 32). The number of patterns
> are:
>
> # of imp = 32           97132  (97132)
> # of imp = 31           26493  (123625)
> # of imp = 30           21460  (145085)
> # of imp = 29           19335  (164420)
> # of imp = 28           18415  (182835)
> # of imp = 27           18703  (201538)
> # of imp = 26           18619  (220157)
> # of imp = 25           19345  (239502)
> # of imp = 24           20390  (259892)
> # of imp = 23           21611  (281503)
> # of imp = 22           22959  (304462)
> # of imp = 21           24675  (329137)
> # of imp = 20           26808  (355945)
> # of imp = 19           29081  (385026)
> # of imp = 18           31938  (416964)
> # of imp = 17           35319  (452283)
> # of imp = 16           39188  (491471)
> # of imp = 15           43899  (535370)
> # of imp = 14           50391  (585761)
> # of imp = 13           57259  (643020)
> # of imp = 12           67062  (710082)
> # of imp = 11           79013  (789095)
> # of imp = 10           95292  (884387)
> # of imp =  9          117109  (1001496)
> # of imp =  8          147810  (1149306)
>
> Depending on the threshold value used (and also the number
> of times the pattern is seen) I can create databases from about
> 100K patterns to 1M patterns, more than that means including
> patterns that are too seldom, their urgency information won't be
> very accurate either due to the small sample size.
>
> Jacques.
>
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/